Harvard Data Science Review’s Founding Editor-in-Chief, Xiao-Li Meng, and Media Feature Editor, Liberty Vittert, interviewed Andrew W. Lo, the Charles E. and Susan T. Harris Professor at MIT’s Sloan School of Management, Director of the MIT Laboratory for Financial Engineering, and a principal investigator at the MIT Computer Science and Artificial Intelligence Laboratory, on the topic of data science for financial markets. Professor Lo was named one of TIME Magazine’s 100 Most Influential People in the world in recognition of his decades of work in financial markets.
In this conversation, they discussed the buzzwords around financial markets: AI, machine learning, big data, and more. They delved into what data can really tell us when it comes to how our money is invested, how machine learning has impacted financial forecasting, to what degree we should rely on AI for investment advice. In particular, Professor Lo shared his insights on whether data can be used to regulate financial markets or turn our individual irrationality rational, and whether there are any data science tools that can help us manage the ups and downs of financial markets, including the risks of cryptocurrencies.
This interview is episode 26 of The Harvard Data Science Review Podcast. This episode released February 23, 2023.
HDSR includes both an audio recording and written transcript of the interview below. The transcript that appears below has been edited for purposes of grammar and clarity with approval from all contributors.
Liberty Vittert: [00:00:05] Hello and welcome to the Harvard Data Science Review Podcast. I’m Liberty Vittert, feature editor of Harvard Data Science Review, and I’m joined by my cohost and editor-in-chief, Xiao-Li Meng. We hear the buzzwords AI, machine learning, big data, etcetera, when it comes to the financial markets, but what can data really tell us when it comes to how our money is invested? Can data be used to regulate the financial markets or turn our individual irrationality rational? Today, to find out this and more, we talk to Professor Andrew W. Lo, the Charles E. and Susan T. Harris Professor of Finance at MIT’s Sloan School of Management. Andrew was named one of Time Magazine’s 100 Most Influential People in the world, with his work on financial markets spanning decades. So, who better to get the investing advice from? Let’s get started.
Xiao-Li Meng: [00:00:57] Well, thank you, Andrew, for joining us today. Over the years, I have heard many, many of your talks and I was incredibly impressed and inspired. I’m quite sure today’s conversation will be the same. I also want to thank you for your incredible support to Harvard Data Science Review; you have written multiple articles, you have helped us to review the articles. I really appreciate it. But today’s topic is really about the financial market because, as you know, this has been on many people’s minds, seeing the market ups and downs. Let’s start with a very basic question from me. The whole stock market has been using statistical data science for many, many years. I guess the New York Stock Exchange started in 1792, and people have probably been doing forecasting since then. Can you give us a brief history of how data have been used in the financial market and what is the current progress there?
Andrew Lo: [00:01:57] Well, first of all, thank you very much for having me on this podcast with you and Liberty. It’s a pleasure to be here. And I would say that ever since financial transactions have existed, there has been data and therefore data science to try to forecast how markets are going to react. If you go back to ancient Babylon and the Mesopotamian era, you’ll see that there are all sorts of records of prices for wheat and other crops. As soon as you start recording prices, people are going to want to forecast them, initially by looking at geometric patterns, using astrology, using bones and tea leaves, and anything else they could dream up. The whole idea of data science is really just a modern version of something that humans have been doing for literally millennia.
Xiao-Li Meng: [00:02:43] Currently, what are the big uses of the data? What are we doing now that is different from this long history?
Andrew Lo: [00:02:51] I would say that there are three changes that have really made a huge difference in what we do now as opposed to what we did many generations ago. The first, of course, is the fact that we have a deeper understanding of how markets work. So, it’s not just looking at various different prices and trying to understand if there are any patterns; it’s being able to come up with economic models to explain those patterns and, therefore, to be able to predict more accurately. The second is that we’re now using computation in ways that we were never able to before because we happen to have the technology to process much larger amounts of data and turn that data into information. And third, is that human behavior has changed. The combination of these three changes has really permanently altered the landscape of human evolution and financial markets in particular.
Liberty Vittert: [00:03:50] It seems that data science helps us understand sort of the volatility of the markets, and I know a lot of your work has been on that. I think about my dad, and as far back as I can remember, there has not been a single week where he has not been stressed or worried about what the market is doing. When everyone talks about right now, for example, being this really uncertain time in the market—is it actually this super uncertain time? How do we measure how volatile a market is? Is this just the random week there are things that feel bad, or is this special? Is this different than anything else?
Andrew Lo: [00:04:26] I think every generation goes through a period where they think that their particular circumstances are unique and so difficult from any other generation’s. I think what we’re seeing now has some commonalities to this phenomenon. Granted, we have not had a pandemic—the kind that we’ve experienced over the last three years—since 1918 with the influenza pandemic. So, it is true that we are living through extraordinary times, and you can actually measure it with specific metrics. For example, it’s not just the volatility of markets—we have many measures for that—but it turns out that over the course of the last three years, the volatility of volatility has actually been quite high. It’s a little bit like acceleration versus velocity. We are really changing the way we measure these kinds of risks. And these new metrics are telling us that we are definitely in very, very uncertain times. If you look at what’s happened over just the last year with the war in Ukraine, oil prices, and the fact that we now have record levels of inflation in the aftermath of the pandemic—those are all things that are relatively unique to our circumstances. But having said that, we also have to remember that even if history doesn’t repeat, it often rhymes; therefore, we have to come up with the kinds of patterns that we’ve seen in the past and make extrapolations. From that perspective, I think it is a very uncertain time, but we see the light at the end of the tunnel, and we’re pretty sure it’s not an oncoming train. We think that there are opportunities for us to normalize over the course of the next 12 to 24 months. So, there’s good reason to be optimistic in the end.
Liberty Vittert: [00:06:13] What are the opportunities to normalize? What’s the path forward over the next 12 to 24 months?
Andrew Lo: [00:06:19] Well, the first path forward is that we are getting out of the pandemic. And by ‘getting out,’ I really mean—
[00:06:26] We shut ourselves down for a two-, two-and-a-half-year period. The best analogy that I can think of is that we put the entire world’s economy into a medically induced coma. When you have a traumatic brain injury, we know that doctors can put you in a medically induced coma to reduce the swelling, and once it gets better, they wake you up. We are now in the process of waking up, and it’s a very difficult process because we are now having to deal with a lot of the supply chain issues that we disrupted over the course of that two- or two-and-a-half-year period. And as a result, inflation has gone through the roof. We’re now managing it, so it looks like it’s coming back down to a normal level. And over the course of the next year or two, we’re going to see a number of jobs being created that really didn’t exist over the last two years. And a number of people going back to work, going back to their daily lives, whatever that means. A new normal, as it were. So, I think that that’s one path forward. The second path forward, we’re going to be creating a number of new industries, and I think that’s going to generate tremendous amounts of growth.
Liberty Vittert: [00:07:37] What are going to be the new jobs? What are these new industries?
Andrew Lo: [00:07:40] Well, a number of the new sectors that have been emerging have to do with technology for dealing with the pandemic. For example, telecommunications: we’re now using Zoom, TikTok, and all sorts of other ways of interacting, social media outlets that we didn’t have before. That’s one example. In biotech, we have now developed new ways of dealing with viruses, namely, mRNA vaccines. But that technology can be applied to many, many other things. So just over the course of the last couple of years, the fact that we’ve had to deal with the pandemic has given us the opportunity to take those technologies and apply them much more broadly. And then, of course, we’ve got other industries like electric vehicles, autonomous driving, fusion energy. Those are areas where technology has made tremendous strides in the last few years, and we’re going to see the consequences of those breakthroughs in the coming decades.
Xiao-Li Meng: [00:08:36] To follow up on that, Andrew, what do you think about specifically how data science is going to help these industries?
Andrew Lo: [00:08:44] Well, I would say that in every single one of those new industries, data science is at the center. For example, take autonomous driving. The only reason that we are able to think about driverless vehicles is because data science has advanced to the point where we can develop machine learning algorithms that can steer a car without any human intervention. But that would not have been possible without the amount of data that we’ve been able to collect and the way of processing the data to generate pattern recognition technology that will allow a vehicle to be steered. Take mRNA vaccines: the way that we can sequence the human genome and other genomes has allowed us to be able to generate vaccines on demand so we can effectively create the equivalent of a 3D printer to generate vaccines whenever we want. That is an amazing technology, and it would not have been possible without data science. So really, every single industry that I can think of that is emerging has emerged because of data science in one way or another. I think that this is the new gold, the new oil of the world today.
Xiao-Li Meng: [00:09:51] Speaking of today, we can’t avoid talking about the FTX fiasco. And I remember—this was probably a year ago—I attended a gathering by your MIT colleague, and you spoke very clearly about the danger or the instability of the cryptocurrency market. I was sitting there thinking that if I had any inclination to want to do something in cryptocurrency then, I would have eliminated those thoughts after listening to your clear warning. Clearly, you anticipated a lot of those things that unfortunately have happened now. How did you assess these risks at the time? Was it based on your experience, your study, your data?
Andrew Lo: [00:10:34] I wish I could claim credit for being able to forecast FTX. I didn’t. I did forecast that there would be some disruptions in the currency markets, in crypto, mainly because it was growing so quickly and people were not doing the kind of due diligence that they normally would because they were so enthused about this market that you knew some kind of accident was waiting to happen. And yes, actually, it’s by looking at past cases of stock market bubbles and financial market panics that you could tell that this is a problem that was developing.
[00:11:08] In particular, in the area of currencies, we actually saw in the 1800s the rise of many kinds of currencies as the American West was being developed. You may have heard of the phrase ‘wildcat banking.’ In those days, banks were started up by a number of individuals that were trying to help the settlers in the West finance their operations, and those banks issued their own form of currency via certificates with various different pictures on them. One of the most common and most widely circulated were certificates with pictures of wildcats on them, because that’s one of the species that populated the West at the time. These certificates were very popular, but, at some point, the banks overextended themselves and the currencies started to fail and people panicked. Eventually, that led to the creation of the Federal Reserve System, because it was pretty clear that the instability of these multiple currencies was creating a real roadblock to progress. So, if you saw what happened in those cases—where this enthusiasm for new currencies led to failures without the proper kinds of regulatory oversight—it’s the same thing with today’s cryptocurrencies.
[00:12:24] And by the way, it’s not just cryptocurrencies, but any new asset class when it first becomes popular and investors overextend and become “irrationally exuberant”—to use a phrase that Bob Shiller1 coined many years ago—you’re going to see problems happening. But I would say that this is an example of progress in all industries, where you go two steps forward, one step back. But there is still progress. So, cryptocurrencies are here to stay. I don’t believe that these are a flash in the pan, but I think that the particular cryptocurrencies that are being traded today may not exist if we see more of these instabilities occurring, and you will have new currencies emerging that will deal with some of these instabilities to the point where, at some point, you may end up seeing sovereigns issuing their own version of the cryptocurrencies. China has already issued it. I suspect that the United States will do that within the next two or three years. We’ll see ‘Fedcoin’ emerge at some point. Once governments start issuing cryptocurrencies, that will create a level of stability that is badly needed in this particular space.
Liberty Vittert: [00:13:37] When you say that in some ways with this cryptocurrency crash, a new market is almost foreseeable, especially if you saw what happened with the banknotes and the creation of the Federal Reserve, was it mainly those who contributed to sort of this crazed investor sense around cryptocurrency? We’re seeing these things where celebrities are leading the general public into investing in cryptocurrency. Or were seasoned investors just as duped by this cryptocurrency craze?
Andrew Lo: [00:14:09] I think it was both, and I think that’s pretty common when you have these types of bubbles. Basically, you have people that are anxious to get into an asset, so they, in some cases, don’t ask the questions that they normally would. They take for granted that something is worth investing in. Maybe they even follow the lead of certain celebrities, as you point out, or other key investors, and they don’t do their own due diligence. They cut corners: that’s the bottom line. But eventually they learn their lessons and unfortunately, they learn it the hard way. And in this case, a number of innocent bystanders can be hurt because it’s their investment dollars that are being managed by these institutional investors. So, I think that’s the real tragedy here, that there are people that can’t afford to lose this money that have lost the money, and that’s when government regulation needs to step in and make sure that this doesn’t happen again.
Liberty Vittert: [00:15:04] That sort of leads me into this question of a theory you came up with. There’s rational market theory, which is that the market is always right—and please correct me if I’m paraphrasing this wrong—but it’s that the market’s always right. If the Dow closes at 10,000 on a Friday and crashes 1,000 points by Monday, then that’s what should happen because the market is rational. You have a theory, the adaptive markets hypothesis, which from my understanding is generally that people mainly behave rationally, but can overreact during periods of heightened market volatility. I guess that’s sort of what you’re saying, is they’re not acting rational when things are coming really fast at them. Is that the case here? What is the reasoning behind your theory, and has it been created by situations like this?
Andrew Lo: [00:15:51] Very much so. The basic idea behind the adaptive markets hypothesis is that financial markets—and, I would argue, economic interactions in general—are much more like a biological ecosystem rather than a physical system that’s subject to immutable laws of motion. We economists, we suffer from a psychological disorder that I call ‘physics envy’; we would love to have three laws that explain 99% of all economic behavior. Instead, we have like 99 laws that explain maybe 3% of behavior. And it’s very frustrating to us. The problem with physics envy is that if we believe that these immutable laws are governing our economy, we then take various different actions, assuming that those never change. But in fact, what we have is a very complex, dynamic ecosystem of different species of investors, brokerage firms, institutional pension funds, and so on. We interact in ways that are much more like species that adapt to changing behavior. In that case, if you look at the world with the lens of adaptive markets, you see that during periods of normalcy, it does make sense that markets are pretty accurate and they are a good measure of value. But every so often we see periods where investors either become irrationally exuberant or irrationally fearful, and in those two extremes, market prices have very little bearing on underlying value. So, I think we need to understand that dynamic, first of all. And then we need to ask ourselves, ‘Are we prepared to deal with those periods of irrational exuberance or fear?’ and if we are, then I think we can deal with a lot of these ups and downs. But most people aren’t aware of them. They take markets as being generally pretty rational, and 90% of the time they’re right. The problem is the10% when markets freak out, that’s a problem for many people, particularly those who are close to retirement and who really can’t afford the losses that, ultimately, they are exposed to.
Xiao-Li Meng: [00:18:00] I want to ask, just from your own study, that I know you use machine learning as well—and thank you for publishing an article about using statistical machine learning to study drug approvals in HDSR—but I want to ask you, just from your personal experience, how much of your research now is influenced by these kinds of machine learning thinking, these pattern recognitions? How much are you relying on these that are hard—I shouldn’t say hard statistical theory, but at least they’re harder than the machine learning theory. What is the data science impact on your personal research?
Andrew Lo: [00:18:36] So, machine learning has been transformative for my research agenda, and it really began years ago when I first applied this to consumer credit problems, trying to understand whether certain consumers would default on their credit card debt and others would not. And I realized that economists had really fallen behind in terms of applying these computational tools, largely because we’d like to understand what our models are telling us. We’d like to develop intuition for why certain predictions are being made. And at the time, these machine learning models were really black boxes. Of course, since then there’s been a lot of development in tools that provide transparency for these machine learning forecasts, so we now can actually figure out what the underlying features are that give rise to these forecasts. Since then, I’ve actually started applying these tools to other problems, and it’s just amazing how effective they are. I think this is the part that really frustrates both economists and, frankly, some statisticians, because statistics has been about forecasting for many hundreds of years, and yet these computational tools that, at first, seem rather simplistic and not particularly sophisticated, they just work really well and it’s very frustrating. But ultimately, I think there has to be some kind of a coming together between statisticians and computer scientists because obviously both deal with data, and as a result, they have a lot to contribute to understand how these tools work.
[00:20:15] So I know that you, Xiao-Li, have been doing a lot of work in this area, and many of your colleagues in statistics have now developed a much deeper understanding of the underlying statistical inference behind machine learning. One of the problems with machine learning early on and the reason that economists rejected that literature at first is because there is no measure of standard errors when you’re making a forecast. Of course, now we do have tools that are able to calculate the accuracy of a machine learning forecast. So, I think the two fields are eventually going to come together in a very productive way. We’re already seeing that now with conferences that feature statistical inference for machine learning, and so on. But I think that economists are probably late to the game. We really need to start developing more and more of an understanding of how our economic models can work collaboratively with these machine learning tools to be able to improve the way that we make our economic forecasts.
Xiao-Li Meng: [00:21:12] Right. I think you’re absolutely right. I think statisticians, we’re probably still going through this period. You know, you start first with denial, and then you gradually accept, and then you join—that process. I have given lots of thought on that, particularly editing Harvard Data Science Review really opened my mind, and so by talking to people like you and others. But there’s always one side of me thinking about that any scientific inquiry requires multiple perspectives. Every discipline has its own responsibility. Probably statisticians do have more of the responsibility about being more critical. Statisticians, as you know, we sometimes joke—probably sometimes not even a joke—that when you torture data, the data will confess. Machine learning is a great way of torturing data. And sometimes a confession is the right confession, but sometimes it’s just made up. So, the question that I have to you, because you’re such a thoughtful leader and I know you have incredible experience in many ways, is how do you guard yourself when you see these patterns to not to be attracted too much, because the pattern’s so attractive and we’re very good at interpreting things into the ones we want? What is the mechanism you use yourself and how do you teach the students to guard against that kind of a tendency when you are using these machine learning algorithms?
Andrew Lo: [00:22:47] Well, I think the first step is to acknowledge that you will never be able to eliminate the problem of overfitting and that there’s always a possibility that you’re fooling yourself. I mean, that’s true with typical statistical estimators as well. Maybe the difference is that statisticians grow up understanding that from day one. We know that there’s always a standard error around our forecast and that we’re liable to overfit the data. I think that the ultimate way that we economists deal with it is to try to develop a deeper understanding of the underlying mechanisms, the economic mechanisms that give rise to these patterns. That’s maybe something that computer scientists don’t spend as much time focusing on simply because of the sheer volume of their data and the applications that they deal with—like image recognition or some other kind of pattern recognition among large pools of data—you don’t really need to worry as much about that underlying motivation. Probably the first example of data science being really effective was the Netflix [Prize] challenge, where they put out all sorts of data about how people chose movies to watch, and they used those algorithms to predict somebody’s preferences. That was an example where you had so much data that it really didn’t matter what the underlying model was. You didn’t understand why it was that somebody enjoyed romantic comedies or slasher movies, you just knew that they did and that you could forecast it. But for economic applications where we don’t have big data—we generally have kind of small data—the overfitting problem is much more serious, so we have to supplement the data analytics with some kind of economic motivation for the underlying patterns. I think that’s really where the tension lies and where the cost and benefit of data science really comes into play. I think we have to supplement with economic models, situations where we don’t have as much data, but areas where we have lots and lots of data, we don’t need as much of the modeling. It’s trying to find that balance in the various different applications that ultimately gives rise to the most successful data science modeling.
Liberty Vittert: [00:25:02] I feel like we’ve talked a lot about what data science has done for statisticians in terms of our ability to model or for economists, but I want to get a little bit to what we can do for the general public or what data science could do for the general public. What feels so difficult to me is how overwhelming the markets can feel, the ups and downs. Going back to my dad, I call them his green or his red days based upon the color of the stocks on his phone, and whether he has a good or bad day based upon that. This noise, these ups and downs, can really consume you, as probably a seasoned investor and as just a general public investor to the point where you really can’t see the—what is it, the trees through the woods, or whatever that saying is. Are there data science tools that can be developed to really help the general public deal with the volatility or the noise of these markets to be able to make good decisions?
Andrew Lo: [00:26:00] Well, there is. There are lots of ways that data science can actually address that issue. And we’re seeing it happening this year, as a matter of fact. So first, let me start by saying that what causes the most stress among investors is the unknown. People have talked about the fact that investors are perfectly willing to take risk, but they really hate uncertainty. Most of us assume that risk and uncertainty are synonyms, but economists actually use them in very different contexts. An economist defines risk as the unknown that you can quantify through probability theory and statistics. But the unknown unknowns, the things that you can’t quantify, they call uncertainty. And so investors are not afraid of risk; they take risk all the time. What investors really freak out about is the unknown unknowns, the uncertainty. So, your dad, like most investors, when they see the stock market going down by 20% and they don’t understand why or how or when it’s going to come back, that’s what causes panic. So how do we deal with that? Well, we deal with it like we deal with most other situations about the unknown: we have to turn the unknown into the known. In particular, the challenge in stock market dynamics is to develop a narrative. So, as humans, we don’t respond to numbers, and we don’t even respond very often to graphs and complicated diagrams. What we respond to are stories, narrative. We need a story. So, the stock market crashed 20% as it did in March of 2020 when the pandemic hit U.S. shores—what was the narrative? I think if we were able to explain to your dad and other investors that the market is down because people are worried about the impact of the pandemic, but we understand what it will do—we’ve seen pandemics before. It will have very devastating consequences to a number of people in the short run, but as an economy, we’re going to come out of this just fine. It’s just going to take a year or two or three, so if you can afford to hold on to your investments, not sell, if you can afford not to panic, then it’ll all be okay. If we had that narrative, I think that that would have reduced market turmoil considerably.
[00:28:23] So how do we do that? Well, financial advisors provide that narrative in some contexts, but not everybody has a financial advisor. The problem with AI is that it hasn’t managed to do that yet. Until this year. With the advent of ChatGPT, we now have an ability to generate narratives, which is extraordinary. I’ve used ChatGPT a bit over the course of the last few weeks, and I’m blown away by what you can do with it, by how responsive it is, by how human its narratives are. So, I believe that over the course of the next few years, we’re going to see dramatic progress in the financial marketplace where we make use of modern AI, particularly natural language processing algorithms that are able to generate not just reasonable responses, but actual narrative that can calm a consumer’s fears about the future. And if that happens, we’re going to see dramatic progress in all sorts of financial products and services.
Xiao-Li Meng: [00:29:25] I think that’s a really interesting insight. I started to play with ChatGPT last night as well, and I was very impressed that I had asked ChatGPT to give me strategies for Harvard Data Science Review to do outreach and fundraising, and it give me six bullet points. I looked through them and I said, ‘Hey, I can implement these.’ That was pretty impressive. Then I asked it to summarize my work, and it came back with a summary that said who I am, which was all fine, and then it said here’s a sample of four of his papers, and none of them I wrote. One of them does not even exist, others were written by other people. So, obviously, there’s something made up. So, at this moment, if there were financial advice given by ChatGPT, how much should the public trust it?
Andrew Lo: [00:30:08] At this point, when it comes to financial advice, I would be extremely cautious about trusting advice from ChatGPT. I don’t think the AI is at a point now where we can really turn over our finances to it. Because, really, this is the first generation of this level of natural language processing, and we’re still working out the bugs. I think that for certain tasks, it’s perfectly fine. For example, summarizing the literature, or producing a shopping list, or getting advice on how to cook a piece of salmon. I think that’s relatively straightforward when it comes to developing the narrative. For things like financial advice, medical advice, personal advice for relationships, I would be extremely cautious about taking ChatGPT at face value because it’s just not there yet. But the fact that it’s improving all the time means that it may not be too long before we can rely on it for that kind of advice. Maybe it’s five years, three years. I don’t know. But I think that the future where we are able to use these AI tools for making very important human decisions, it’s now reachable, I think, in our lifetimes.
Xiao-Li Meng: [00:31:28] By saying that, you are implying that human intelligence at some point is probably reachable. I don’t know if you’re implying that, but it sounds like you are very optimistic about it. The current ChatGPT does give a sense of what can come, but there’s also the fundamental question: Are these just low-hanging fruit, so the machine can do it? Are there some limits by this kind of pattern recognition without really understanding it? How close can we get to the human intelligence? Or maybe even exceed it, right? Because ultimately, we hope the machine can do things that we cannot do.
Andrew Lo: [00:32:01] Well, that’s a very deep question, and I’m not sure I’m qualified to answer it. I’m certainly not a neuroscientist or a philosopher, but I can tell you from my own perspective as a researcher in economics and having written about intelligence from an evolutionary perspective with the mathematical models that I developed, I have a very specific view of what intelligence is. In my view, I can answer the question that you asked. But first, let me describe what that view is. So, for me and for the mathematical models that I developed, intelligence is really captured as an adaptation that provides you with an advantage for survival. That’s it, plain and simple. Whether it’s the opposable thumb or the ability to solve differential equations, the idea behind intelligence in all cases, from my perspective, is some type of human adaptation that allows us to increase our chances for survival. If you agree that that is the definition of intelligence, then I would argue that AI is already here in terms of being able to display a certain degree of intelligence. As it evolves, as AI becomes more sophisticated, it will at some point achieve the same or greater level of intelligence as humans.
[00:33:25] In fact, you could argue that the idea of self-awareness, sentience, is a form of intelligence that comes with a certain degree of processing power. At some point, AI may become self-aware simply because of the ability to process and detect patterns in various different guises and contexts. You know, the human brain has about 100 billion neurons. That’s a big number. But that’s a finite number, and it’s a number that can be replicated in silico. And the human being—as unique as we are—is a product of evolution. Who’s to say that we won’t get the same level of evolutionary adaptation from our machines? So more and more, as scientists study the nature of human thought and how the brain works, it seems that what we are able to do is not so unique, but it’s really computation at a certain level of sophistication. So, in other words, human cognition is really a matter of degree, rather than something that is just totally unique. And so from my perspective, based upon what I imagine intelligence is, I think we’re already there to a certain degree, and we’re going to see exponential progress over the course of the next decade or two in various different types of AI that, at some point, will become self-aware, and that we will then have a whole new field that we have to deal with, which is the ethics of AI and how we as humans relate to the machines.
Liberty Vittert: [00:35:07] You talk about how there’s the chance that it could develop this awareness, and I can’t help but look at the ChatGPT issues of bias, where we’ve seen in all algorithms that bias has been one of the biggest issues, I think, with machine learning. For example, if you asked ChatGPT to write a paragraph about why fossil fuels will be good for human happiness, it says, ‘I’m not able to do that because I don’t believe in fossil fuels. They’re bad for climate change and I refuse to do that. It goes against my ethics.’ Which is clearly a person putting that in, not the machine choosing that. So, will these ChatGPTs be able to get the awareness to understand their own bias, to overcome the bias issues that we’ve seen over and over again permutate machine learning?
Andrew Lo: [00:35:56] Well, that’s another deep question. And given that humans can’t even see their own biases in many cases, I don’t know that ChatGPT is going to be any different. I think it’ll be a challenge. One of the differences, though, in how we can approach it is since we are designing ChatGPT, we have the ability to incorporate some form of bias detection and bias correction at the very outset. The fact that we are aware of the existence of bias means that we can actually do something to prevent it. I think it’s a very positive trend. There’s been talk, for example, over the last few years about ‘Asian hate.’ And I have to say that when I grew up in New York in the 1960s and 1970s, I experienced Asian hate as well, but we didn’t call it Asian hate at the time; we called it Tuesday. Saturday. Friday. And all the other days. But the fact that we are giving it a name means that we can actually start to manage it. I think we are making progress as a society, and the fact that we are building these systems now means that we have a chance—doesn’t mean it’ll necessarily happen—but we do have a chance to incorporate some of these bias detection and bias correction algorithms at the very outset. We may be able to develop intelligence that, ultimately, is fairer than we are and may be wiser than we are.
Xiao-Li Meng: [00:37:19] I would imagine that the problem is that humans have bias because, as you said, each one of us only has this many neurons. We only have this way of thinking. The ChatGPT or any of these models, these algorithms, supposedly will be able to kind of synthesize all kinds of opinions. So really, accumulated, it could possibly be better than any single individual trying to correct the bias, because whenever trying to correct bias, we inject our own bias.
Andrew Lo: [00:37:50] Well, the ultimate goal from a societal perspective is to do the greatest good for the greatest number. I think that we all agree that this kind of utilitarian perspective pervades much of our public policy and our laws and our general actions. The problem is that the greatest good for the greatest number doesn’t necessarily guarantee that everybody is going to be brought along and favored in that way. The greatest good for the greatest number may mean that there is a significant subset that will be marginalized, and there’s just not a whole lot we can do about it. The bottom line is that because resources are finite, we are dealing with a zero-sum game in human society, and that means that there are going to be winners and losers. I think the idea behind bias, though, is that all of us want a fair shot at being able to win as opposed to automatically being predetermined to lose. So, I think the greatest good for the greatest number ultimately has to have a statistical angle to it, which is that there has to be some mechanism by which all of us are given an equal probability of succeeding in certain respects. And I don’t know if that’s possible.
Xiao-Li Meng: [00:39:23] I’m quite sure that statisticians are very happy to hear what you just said, because this goes back to the most basic statistical sampling scheme: simple random sampling, to ensure everybody has an equal chance of being selected even though most people will not be selected because it’s a small sample. But it’s the notion of being included with that equal chance.
[00:39:44] At the end of every episode, we always talk about the magical wand. The question is that if you could give everyone one piece of investment advice they should follow, and if you can wave the magic wand, what would that be?
Andrew Lo: [00:39:59] Well, I think the one piece of advice is that there does not exist one piece of advice for financial management. One size does not fit all, and you should be wary of people that are trying to sell you that kind of advice. I think the more sophisticated answer is that financial markets are highly dynamic and you need to be equally dynamic in how you respond to them. That means educating yourself about different financial options and keeping track of the amount of fees you’re paying. Low-cost mutual funds are better than high-cost mutual funds, on average. Diversify your portfolio and spend more time learning about financial markets. The world we live in today is a lot more complicated than 20 years ago. That’s true with virtually every aspect of our lives, whether it’s technology or health, and certainly, financial markets are no different. In the same way that the old advice for your diet of ‘eat to your heart’s content’ is no longer the best advice, and you have to worry about carbs and the amount of sugar you’re taking in. That same set of sophisticated situations has arisen in financial markets as well, so you need to educate yourself in the various different ways you can invest your money and protect your assets for the future.
Liberty Vittert: [00:41:27] Man, I was looking for a, like, ‘Buy low, sell high’ kind of thing here, but I like this one. This is more thoughtful. [Laughing.]
Andrew Lo: [00:41:35] That works, too.
Xiao-Li Meng: [00:41:37] Thank you so much for this absolutely wonderful conversation. Thank you.
Andrew Lo: [00:41:41] Thank you. It’s been a pleasure and an honor. The Harvard Data Science Review is a fantastic outlet for all things data science, and I continue to learn from every issue. So, thank you for having me on.
Liberty Vittert: [00:41:57] Thank you for listening to this week’s episode of the Harvard Data Science Review Podcast. To stay updated with all things HDSR, you can visit our website at hdsr.mitpress.mit.edu or follow us on Twitter and Instagram @theHDSR. A special thanks to our producers, Rebecca McLeod and Tina Tobey Mack, and assistant producer Arianwyn Frank. If you like this podcast, don’t forget to leave us a review on Spotify, Apple, Podbean, or wherever you get your podcasts. This has been the Harvard Data Science Review Podcast: everything data science and data science for everyone. Thanks so much for listening.
Andrew W. Lo, Xiao-Li Meng, and Liberty Vittert have no financial or non-financial disclosures to share for this interview
©2023 Andrew W. Lo, Xiao-Li Meng, and Liberty Vittert. This interview is licensed under a Creative Commons Attribution (CC BY 4.0) International license, except where otherwise indicated with respect to particular material included in the interview.